NVIDIA GPUs by the hour. From one card to a full cluster.
Train, fine-tune and serve AI models on the latest NVIDIA hardware. Pay on-demand, save with spot, or reserve for up to 28% off, and see every price up front.
See every GPU in one table
Prices per GPU per hour, excluding taxes. vCPU and RAM are per GPU. Spot capacity may be interrupted.
Built for serious AI work.
Bare metal or virtual
Get the whole server with no virtualisation overhead, or a VM that starts in minutes.
Cluster networking
High-speed InfiniBand or RoCE links between nodes for multi-node training.
Ready-to-use images
NVIDIA drivers, CUDA, PyTorch and common AI tools come pre-installed.
Kubernetes or Slurm
Managed clusters with your choice of scheduler, set up by our team.
Fast shared storage
A shared file system and S3-compatible object storage sit next to your GPUs.
Three ways to pay
On-demand by the hour, cheaper spot capacity, or reserved terms of 1, 6 or 12 months.
Which GPU do you need?
Not sure? Here is a quick guide. Or tell us about your model and we'll recommend the cheapest GPU that does the job.
Ask for a recommendation- L4 / L40S: running AI apps (inference), image and video generation, and smaller models on a budget.
- RTX PRO 6000: 96 GB of memory for inference, fine-tuning and 3D or graphics work.
- A100 / H100: the standard for training and fine-tuning. Mature and widely supported.
- H200: like the H100 with more memory, for large language models.
- B200 / B300: the latest Blackwell GPUs, for the fastest training and high-volume inference.
- GB200 / GB300 NVL72: 72 GPUs working as one, for frontier-scale training. Quoted per project.
Need 64 GPUs or more?
Large clusters get custom pricing and dedicated capacity. Tell us your size, timeline and region.